silence_periods() error when using mode='noise'

[edmundchong]

Hi, I've been trying to use spikeinterface.preprocessing.silence_periods() which works fine when mode='zeros'. However, when I try mode='noise',

1287 dtype = np.dtype(dtype).name # Cast to string for serialization
1288 if dtype not in ("float32", "float64"):
-> 1289 raise ValueError(f"'dtype' must be 'float32' or 'float64' but is {dtype}")
1290 assert strategy in ("tile_pregenerated", "on_the_fly"), "'strategy' must be 'tile_pregenerated' or 'on_the_fly'"
1292 if np.isscalar(noise_levels):

ValueError: 'dtype' must be 'float32' or 'float64' but is int16

My rec.dtype returns int16 indeed. Do I need to do some conversion before I can use this? Thanks!

[chrishalcrow]

Hi @edmundchong , yes at the moment silence_periods with mode='noise' requires the recording to be float type. You can make this happen by doing something like:

import spikeinterface.full as si
recording_with_dtype_float = si.astype(recording, dtype='float64')

Before you do the silence_periods step.

[edmundchong]

thank you, this works! but @chrishalcrow does this change of dtype have any other functional consequences downstream? I changed it back to int16 anyway after the silence periods just in case. Thanks!

[chrishalcrow]

From an accuracy point of view, changing from int to float should only improve things.
But the recording will take up more space (if you save it) and more RAM on compute. So it might make some things slower. Nowadays, CPUs are pretty well optimized for floats, so there usually isn't a huge difference in compute times - try it out!